Cheng, Yu-TingYu-TingChengLiu, Ming-JieMing-JieLiuHuang, Yu-TangYu-TangHuangLIANG-CHIA CHEN2026-01-152026-01-152026-02-2402632241https://www.scopus.com/record/display.uri?eid=2-s2.0-105024719609&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/735343Optical measurement is widely used for non-destructive inspection, but the continual scaling of semiconductor structures has exceeded the limits of conventional visible-light imaging. This study develops a deep-learning resolution-enhanced imaging approach grounded in physical optics. A novel Point Spread Function (PSF) calibration target, optimized for bright-field microscopy, enables consistent PSF modeling and high-quality imaging. A Fourier optics-based imaging model generates synthetic training data, combined with experimental data, to train a PSF-based Multi-channel Deconvolution U-Net (PSF-MDUnet) that reconstructs high-frequency details. Experiments show a nearly 50 % reduction in measurement bias for micro-scale structures, achieving an average measured bias of 0.86 % and a 55 % increase in depth of field within a 1-pixel tolerance. In off-axis regions, the radius measurement error decreases from 28.9 nm to 4.6 nm. The system surpasses the Rayleigh limit (419 nm), resolving features down to 329 nm, achieving a 1.28× improvement in lateral resolution and 115 % contrast enhancement. This advancement significantly improves critical dimension measurements for advanced semiconductor packaging.falseCritical dimension metrologyImage resolution enhancementPSF calibrationPSF-based Multi-channel Deconvolution U-NetSemiconductor metrologySemiconductors[SDGs]SDG7journal article10.1016/j.measurement.2025.1200392-s2.0-105024719609